How to tell if an image is AI-generated: what to check and what no longer works

There’s no foolproof way to tell if an image is AI-generated. What works is stacking three checks: look in the file for a label or credential saying it was made with AI, run it through the big generators’ invisible-watermark detectors and trace its origin with a reverse image search. Looking at hands, text and shadows helps, but it’s no longer enough.

A credential signed by the generator is strong evidence; a missing label proves nothing. We checked in October 2026: after re-compressing two images made with ChatGPT using default settings, neither kept its label.

Can you tell for sure if an image is AI-generated?

Rarely with a single method. If the file carries a credential signed by the generator, or a detector finds its watermark, you have strong evidence. If it carries nothing, you know nothing: metadata gets lost along the way, and each watermark can only be read by its own technology’s detector.

What are Content Credentials and the IPTC label?

They’re data that travel inside the file and describe how the image was made. Content Credentials follow the open C2PA standard: a signed record of the tool that created the image, the date and any edits. The IPTC label is simpler: a field that says, for example, “Created using Generative AI.”

The coalition behind the standard, C2PA, includes Adobe, Google, Meta, Microsoft and OpenAI, among others. The record stores a cryptographic fingerprint of the image: if anyone alters it, the two no longer match, as the coalition explains in its technical explainer. And it isn’t just about AI: compatible cameras use the same standard to certify real photos.

The IPTC label comes from the Digital Source Type vocabulary of the IPTC, the news media’s standards body. Three of its values:

• trainedAlgorithmicMedia: created using generative AI.
• compositeWithTrainedAlgorithmicMedia: edited using generative AI, for example to fill in or extend a photo.
• digitalCapture: captured from real life with a digital camera.

How to check Content Credentials

1. Get the original file: the direct download, not a screenshot or a forwarded copy.
2. Open the public Content Credentials verifier and choose the image from your device or drag it onto the page.
3. If it has credentials, check its history; if nothing shows up, don’t conclude anything: the verifier itself warns that a lot of content doesn’t carry that information yet.

What happens to the AI label when you compress an image?

It usually gets lost. We tested it in October 2026 with two ChatGPT images: one downloaded straight from ChatGPT and the cover of our article on domains, hosting and DNS. When we re-saved them the way a compressor would, both lost their label; only the cover’s survived when we explicitly asked to keep the metadata.

The first, generated in June 2026, carried a 28.4 KB C2PA manifest signed with a certificate issued to OpenAI. It said the image was created by “gpt-image” 2.0 (value trainedAlgorithmicMedia) and that an invisible watermark had been added. When we converted it to WebP with sharp, a free image-processing library, the manifest disappeared, even when we asked it to keep the metadata.

The cover no longer has a manifest: it got lost at some step before we published it. It carries a 728-byte XMP block that we added ourselves, with the IPTC value trainedAlgorithmicMedia and a credit line (in Spanish) for an illustration made with AI using ChatGPT. When we redrew it on a Chrome canvas, as many tools that compress inside the browser do, that block was wiped too.

Images that come out of our image compressor lose that metadata too, because they’re also redrawn on a browser canvas. The takeaway: if the metadata is there, it helps; if it isn’t, it proves nothing.

What is an invisible watermark, and who can read it?

It’s a signal hidden in the pixels themselves: the eye can’t see it, and it doesn’t go away when the metadata is wiped. Google DeepMind designed its own, SynthID, to withstand cropping, filters and compression. The catch: only that technology’s detector recognizes it, so it says nothing about images from other generators.

This is how OpenAI, Google and Meta mark or label images today:

How do you find out where an image came from?

With a reverse image search: the search engine shows you where else the image appears and since when. It won’t tell you how it was made, but it will tell you if that “photo from today” was already circulating years ago, if only anonymous accounts post it or if a known news outlet published it first.

On Google, the About this image feature shows when Google first saw it, which pages use it and, if the file carries metadata, whether it was taken with a camera or made with AI. It’s only available in some regions, and Google warns that this metadata can be modified with editing tools.

Do visual clues still work?

As a hint, yes; as proof, no. In a Microsoft experiment published in 2025, with about 287,000 answers from more than 12,500 people, participants told real images from AI ones correctly 62% of the time, only slightly better than chance. They did worse with landscapes than with portraits.

If you want to inspect an image, look for:

• Hands and fingers that merge, multiply or go missing.
• Text on signs, labels or screens that says nothing.
• Shadows and reflections that don’t match the light.
• Backgrounds with melted, repeated or half-finished objects.

Is it illegal to post AI-generated images?

Not in itself: European law asks for transparency. Since August 2, 2026, Article 50 of the EU AI Act has required anyone offering a generator to mark its images in a machine-readable format. And anyone who uses AI to create a deep fake, except for purely personal use, must disclose that it’s artificial.

Under the Act, a deep fake is AI-generated or manipulated image, audio or video content that resembles real people, objects, places or events and would falsely appear to be authentic. For evidently artistic, satirical or fictional works, a disclosure that doesn’t spoil the work is enough. And generators already on the market before that date have until December 2, 2026 to mark their images, under an amendment adopted in July 2026.

Outside the European Union, the rules depend on each country.

What changes for you and your business?

• Don’t share without checking. An image that arrives with a sense of urgency, like an unbelievable deal or outrageous news, deserves the minute these checks take.
• Be wary of photos used as proof. If a customer complains with a photo of a damaged product or someone sends you a payment receipt, ask for the original file and compare it with another source.
• Label what you make with AI. A visible caption, like the “Illustration made with AI (ChatGPT)” credit on the cover we tested, reaches everyone even if the file loses its metadata.
• Keep the original. Hold on to the file the generator gave you, with its credentials, and check that your editing app doesn’t strip the label before you publish.

What gets exaggerated about spotting AI images

• “Detectors are always right.” Not even OpenAI’s: in 2024, its classifier recognized about 98% of images from its own DALL·E 3 in internal tests, but flagged only 5% to 10% of images made by other generators.
• “No metadata means it’s real.” A screenshot or a compressor wipes it, as our test showed, and OpenAI acknowledges it also gets lost through uploads and downloads. Its absence says nothing.
• “AI always gets hands and text wrong.” That’s no longer a rule: when OpenAI launched GPT-4o image generation in March 2025, it highlighted that it renders text accurately, while admitting limitations with small text and multilingual text.

What should you do when you doubt an image?

1. Look for a visible caption or the social network’s AI label.
2. Run the original file through the Content Credentials verifier.
3. Try Google’s and OpenAI’s detectors.
4. Do a reverse search to see where and when it appeared.

If you still can’t tell after all that, treat it as doubtful: don’t share it as true or use it as proof.

And if you work with images, Vertize Plus’s free tools let you compress them or remove the background with AI without them ever leaving your browser.

Frequently asked questions